Rethinking Real-World MRI Denoising: Learning from Physical Noise

Conference: ECCV 2026
Paper: ECCV Official
Code: https://github.com/Deep-MI/YADO
Area: Medical Imaging
Keywords: MRI Denoising / Physical Noise2Noise / Diffusion Re-noising / Cross-Contrast Guidance / Self-Supervised Learning

TL;DR

Addressing the core obstacle in magnetic resonance imaging (MRI) where noise-free reference targets are physically unobtainable, this paper introduces YADO, which directly leverages repeated scans to establish physical Noise2Noise (pN2N), generalizes to single acquisitions via diffusion-based re-noising (ReN2N), and incorporates co-acquired contrast guidance to outperform 17 baselines across 14 benchmarks while matching brute-force physically averaged scans.

Background & Motivation

Magnetic resonance imaging (MRI) is a cornerstone of modern neuroimaging and clinical diagnostics, yet its physical acquisition process is inherently plagued by measurement noise and spatial correlations arising from coil arrays and accelerated reconstruction. This acquisition noise blurs tissue interfaces and diminishes the reliability of downstream biomarkers, particularly for subtle, patient-specific pathologies such as perivascular spaces (PVS) and white matter hyperintensities (WMH). In medical MRI, completely noise-free ground truth images are physically unobtainable (\(\bar{\mathbf{n}} \neq 0\)). Consequently, supervised training with clean targets (Noise2Clean) remains impossible, compelling the field to rely on surrogate self-supervised objectives.

Existing self-supervised denoising paradigms, however, suffer from structural degradation and severe domain shifts. Blind-spot approaches such as Noise2Void and Noise2Self operate under the strict assumption of voxel-wise noise independence—an assumption routinely violated by spatially correlated MRI noise; enlarging blind-spot masks drastically compromises high-frequency anatomical detail. Conversely, methods following the Noisier2Noise paradigm inject synthetic white Gaussian or Rician noise into already corrupted images to train models to reverse incremental degradation. This artificially shifts the input distribution toward higher noise regimes and induces domain discrepancies that lead to incomplete noise suppression or residual artifacts at inference. Zero-shot denoisers (such as Deep Image Prior) escape paired requirements but impose prohibitive per-scan optimization costs and remain highly vulnerable to overfitting specific noise realizations.

To resolve this persistent compromise, this work revisits the theoretical foundation of Noise2Noise within physical imaging constraints. Large-scale neuroimaging studies routinely collect independent in-session or same-day repeat scans during quality assurance protocols, and standard clinical workflows almost universally acquire multi-contrast sequences (such as T1w, T2w, and FLAIR). The core idea is: directly leverage physically independent repeated acquisitions to formulate an unbiased physical Noise2Noise (pN2N) training target, synthesize noise-decorrelated yet anatomically faithful pairs for unrepeated scans using a diffusion-based re-noiser (ReN2N), and guide denoising with co-acquired multi-modal contrasts to establish an authentic physical MRI denoising framework, YADO.

Method

Overall Architecture

The YADO (You Accurately Denoise real Observations) framework recovers high-fidelity anatomical morphology directly from real-world noisy MRI observations. The pipeline integrates three primary mechanisms: input slices corrupted by physical acquisition noise are co-registered and concatenated with auxiliary contrast guidance, then passed into a high-capacity 2.5D multi-plane U-Net backbone. During training, datasets with physical repeats follow the physical Noise2Noise (pN2N) formulation; for single-acquisition datasets, a conditional diffusion re-noiser trained on real repeats synthesizes decorrelated noisy pairs (ReN2N). During inference, the network takes the target scan alongside available co-acquired contrasts, performing multi-planar inference to produce noise-suppressed anatomical reconstructions.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Noisy Real-World MRI Input<br/>(Single or repeated scan)"] --> B["Physical Repeat Pair Learning (pN2N)<br/>Exploit independent physical repeats"]
    A --> C["Diffusion Re-noising Augmentation (ReN2N)<br/>Conditional DM synthesizes decorrelated pairs"]
    D["Co-acquired Multi-contrast MRI<br/>(T2w / FLAIR guidance)"] --> E["Cross-Modality Guidance (gYADO)<br/>Inject complementary anatomical priors"]
    B --> F["2.5D Multi-Plane U-Net Backbone<br/>(5-slice slab with axial/coronal/sagittal fusion)"]
    C --> F
    E --> F
    F --> G["High-Fidelity Denoised Output<br/>(Matches NEx=7 physical averaging)"]

Key Designs

1. Physical Repeat Pair Learning (pN2N): Eliminating domain shifts and information loss at the source Existing self-supervised approaches compromise performance by artificially masking context or adding synthetic noise distributions that do not reflect scanner physics. pN2N overcomes this by capitalizing on independent physical repeat acquisitions \(X_1 = X' + \mathbf{n}_1\) and \(X_2 = X' + \mathbf{n}_2\) collected from the same participant within the same scanning session. Because acquisition thermal noise and non-systematic distortions across repeat scans are statistically independent (\(\mathbf{n}_1 \perp \mathbf{n}_2\) with \(\mathbb{E}[\mathbf{n}]=0\)), training a regression network \(f_\theta\) to map \(X_1\) directly to \(X_2\): $\(\arg\min_\theta \mathbb{E}_{X_1, X_2} \left[ \|f_\theta(X_1) - X_2\|_2^2 \right]\)$ yields the conditional expectation of the clean signal \(\mathbb{E}[X'|X_1]\) under Noise2Noise theory. Unlike blind-spot or degradation-based training, pN2N preserves 100% of the input anatomical context and requires no artificial corruption, preventing domain shifts between training and real-world deployment.

2. Diffusion Re-noising Augmentation (ReN2N): Generalizing pair-based supervision to unrepeated scans While massive neuroimaging consortia preserve physical repeats, typical clinical archives contain only single-acquisition scans, precluding direct pN2N supervision. To bridge this gap, ReN2N trains a conditional denoising diffusion probabilistic model (DM) on available repeat cohorts using a continuous square-root cosine (\(\sqrt{\cos}\)) noise schedule and a \(v\)-prediction objective. Unlike naive Gaussian or Rician noise addition, the reverse diffusion process generates a latent clean signal estimate \(\hat{X}'\) before terminating with realistic, spatially correlated physical noise: $\(\hat{X}_t^{(\text{DM})} = \hat{X}'^{(\text{DM})} + \hat{\mathbf{n}}_t^{(\text{DM})}\)$ By sampling four stochastic re-noised instances per single input and combining them with the original scan into a five-instance pool, ReN2N produces synthetic pairs with decorrelated noise while perfectly preserving patient-specific anatomy. This design extends N2N training to arbitrary single-noisy datasets and serves as a powerful regularizer that markedly improves cross-site generalization.

3. Cross-Modality Guidance (gYADO): Tightening the posterior with orthogonal anatomical contrast Clinical neuroimaging protocols routinely acquire multiple pulse sequences (such as T1-weighted, T2-weighted, and FLAIR) in a single patient visit. Different pulse sequences possess orthogonal sensitivity to tissue relaxation properties: anatomical boundaries or subtle microvascular structures that exhibit faint contrast in T1w may appear sharply demarcated in T2w. gYADO exploits this synergy by aligning auxiliary contrasts via rigid registration and concatenating them as multi-channel inputs to form \(f_\theta(X, C) \approx \mathbb{E}[X'|X, C]\). Incorporating complementary anatomical priors directly tightens the posterior distribution, preventing the denoiser from over-smoothing low-contrast structures such as perivascular spaces.

Loss & Training

YADO employs a 2.5D deep convolutional architecture based on a five-stage ResNet U-Net ([256, 256, 512, 512, 1024] channels) equipped with multi-head self-attention ([0, 0, 2, 2, 4] heads), totaling 213 million parameters (213M). To reconcile GPU memory efficiency with cross-slice continuity, the model operates on five-slice slabs, sampling randomly across axial, coronal, and sagittal orthogonal planes during training, and averaging predictions across all three orientations at test time.

The network is optimized using an \(L_2\) regression loss. To prevent zero-information background air voxels from biasing the gradient dynamics away from parenchymal boundaries, a brain mask \(M_{\text{brain}}\) extracted via FastSurfer is applied, down-weighting extracranial loss contributions by \(\varepsilon = 0.01\): $\(\mathcal{L}(\theta) = \| (X_2 - f_\theta(X_1)) \odot (M_{\text{brain}} + \varepsilon (1 - M_{\text{brain}})) \|_2^2\)$ Training is executed with the AdamW optimizer at an initial learning rate of \(10^{-4}\) and a batch size of 96. Model weights are tracked with an exponential moving average (EMA, \(\beta = 0.999\)), using the final EMA checkpoint directly to guarantee stable convergence without requiring a dedicated validation set.

Key Experimental Results

Main Results

On representative T1-weighted brain MRI benchmarks spanning OASIS-3, CHDI (Huntington's disease cohort), Kirby, HCP-Epilepsy, HCP-Young Adults, and the Rhineland Study (RS)—covering three scanner vendors (Siemens, Philips, GE) and 0.7–1.0 mm isotropic/anisotropic resolutions—YADO was benchmarked against 17 competing methods. Evaluation metrics include Structural Similarity (SSIM, %) computed against independent physical repeat reference scans and absolute contrast-to-noise ratio for perivascular spaces (PVS CNR).

Method	OASIS-3 (1mm) SSIM	CHDI (1mm) SSIM	Kirby (1mm) SSIM	HCP-E (0.8mm) SSIM	RS (0.8mm) SSIM	RS PVS CNR	Overall Mean Rank (all rank ↓)
gYADO (ReN2N)	89.77 ± 4.82	94.59 ± 1.47	(95.63 ± 0.78)	90.91 ± 3.24	94.59 ± 2.25	2.85 ± 1.26	1.46
uYADO (ReN2N)	89.41 ± 4.87	94.33 ± 1.59	95.99 ± 0.90	90.60 ± 3.38	94.37 ± 2.39	2.74 ± 0.75	2.09
gYADO (pN2N)	90.03 ± 4.35	94.11 ± 1.42	(94.63 ± 0.76)	89.96 ± 3.31	94.31 ± 1.59	2.85 ± 1.66	4.17
uYADO (pN2N)	89.39 ± 4.69	94.25 ± 1.54	95.69 ± 0.85	89.65 ± 3.82	93.83 ± 2.23	2.63 ± 0.81	4.55
S-N2Void (Blind-spot)	88.69 ± 5.05	93.67 ± 1.74	95.59 ± 1.01	88.43 ± 3.73	92.62 ± 2.51	2.17 ± 0.65	7.81
BM4D (Transform filter)	88.28 ± 4.52	93.40 ± 1.67	95.33 ± 0.90	89.37 ± 3.66	92.93 ± 2.43	2.13 ± 0.83	7.96
N2Void	88.60 ± 5.20	93.78 ± 1.76	95.72 ± 0.99	88.19 ± 3.86	92.63 ± 2.63	2.24 ± 0.62	8.58
Noisier2Noise	88.59 ± 4.82	93.53 ± 1.71	95.58 ± 0.90	89.47 ± 3.28	93.42 ± 2.60	2.28 ± 0.83	9.26
ANTs NLM (Adaptive NLM)	87.89 ± 4.58	93.11 ± 1.69	94.97 ± 0.91	88.95 ± 3.75	92.62 ± 2.64	2.36 ± 0.66	9.49
Neigh2Neigh	87.94 ± 4.31	93.03 ± 1.53	94.59 ± 0.76	89.08 ± 3.27	92.68 ± 2.31	2.31 ± 0.89	9.75
DIP (Zero-shot)	87.87 ± 5.34	93.27 ± 1.80	95.31 ± 1.16	89.04 ± 4.09	92.88 ± 3.07	2.33 ± 0.60	9.84
RED-WGAN	86.35 ± 5.34	91.38 ± 1.95	93.42 ± 1.13	87.21 ± 3.50	90.49 ± 3.19	2.18 ± 0.48	15.35
Identity (Undenoised input)	86.91 ± 5.65	92.66 ± 2.09	95.03 ± 1.21	87.09 ± 4.46	91.60 ± 3.25	2.33 ± 0.46	14.35

Ablation Study

To assess cross-contrast generalization, the framework was evaluated on Rhineland Study (RS, 0.8 mm) T2-weighted and FLAIR acquisitions alongside downstream tasks, including white matter hyperintensity (WMH) CNR and automatic lesion detection F1 score (%) evaluated using the SHIVA segmentation pipeline.

Model / Configuration	RS T2w (0.8mm) SSIM	RS T2w PVS CNR	RS FLAIR (0.8mm) SSIM	RS FLAIR WMH CNR	SHIVA WMH Detection F1 (%)	Global Average Rank (rank ↓)
gYADO (pN2N)	89.65 ± 8.60	3.12 ± 0.87	89.60 ± 5.07	3.40 ± 0.45	64.23 ± 19.13	3.55
gYADO (ReN2N)	88.98 ± 9.69	2.91 ± 0.85	89.28 ± 5.49	3.33 ± 0.41	62.42 ± 21.01	3.63
uYADO (ReN2N)	88.58 ± 10.43	2.59 ± 0.99	88.25 ± 6.33	3.08 ± 0.33	61.02 ± 21.33	4.60
uYADO (pN2N)	88.80 ± 10.44	2.69 ± 0.95	88.36 ± 6.30	3.14 ± 0.39	61.58 ± 21.53	5.60
BM4D	87.28 ± 10.81	2.52 ± 3.18	87.76 ± 6.64	3.21 ± 0.33	62.83 ± 20.32	4.74
S-N2V (Blind-spot)	87.69 ± 11.35	2.38 ± 0.95	86.60 ± 7.02	3.36 ± 0.34	62.75 ± 20.17	5.52
ANTs NLM	86.76 ± 10.47	2.32 ± 0.80	87.20 ± 6.38	3.37 ± 0.33	62.89 ± 19.96	6.00
Ner2N (Noisier2N)	87.00 ± 11.66	2.21 ± 0.78	85.78 ± 7.05	3.34 ± 0.32	62.82 ± 19.95	7.65
Identity (Undenoised raw)	86.99 ± 11.67	2.22 ± 0.79	85.73 ± 7.06	3.37 ± 0.34	62.45 ± 20.19	8.41

Key Findings

• The real-world deep denoising pitfall: Among the 17 evaluated baseline approaches, 7 methods ranked lower than the trivial "Identity" baseline (which leaves noisy images unedited). Because physical acquisition noise possesses spatial correlations and real reference scans carry unavoidable aleatoric noise (\(c_{\mathbf{n}_{\text{ref}}}\)), synthetic degradation models underfit physical noise and inadvertently blur real anatomy or hallucinate high-frequency artifacts. The classical BM4D filter outperformed heavy deep learning baselines such as RED-WGAN, BME-X, and DDM2.
• Superior generalization of ReN2N: Across 22 independent T1w generalization experiments, ReN2N outperformed physical repeat training (pN2N) in 20 instances. This confirms that diffusion-synthesized noise pairs not only unlock pair-based supervision for unrepeated archives but also act as a potent physical data augmentation mechanism that mitigates overfitting to scanner-specific noise traits.
• Cross-contrast guidance boundaries: Co-acquired guidance provides consistent improvements in SNR and anatomical clarity (raising PVS CNR from 2.63 to 2.85). However, when the guidance sequence suffers from severe anisotropy or low slice resolution (as observed with Kirby T2w scans), the anisotropic blurring spills into the target volume, indicating that guidance sequences must maintain compatible spatial resolution.

Highlights & Insights

• Rooted in physical reality over surrogate synthetic shortcuts: Exposes the systemic flaws of conventional self-supervised MRI denoising that rely on artificial Gaussian/Rician noise or heavy masking. By formulating pN2N on physical repeat acquisitions, the approach recovers clean anatomical signals with an unbiased regression loss.
• Diffusion models repurposed as physical noise generators rather than denoisers: Overcomes the common pitfalls of end-to-end generative models in medical imaging (hallucinations, excessive inference latency, loss of fidelity). It confines the stochastic power of diffusion models to offline physical noise synthesis, preserving deterministic, real-time 2.5D U-Net inference.
• Routines scans matching brute-force physical averaging: A single routine clinical scan (NEx=1, taking a few minutes) processed with YADO achieves structural detail and signal-to-noise ratios matching physical repeat averaging across 7 independent acquisitions (NEx=7, acquisition time exceeding 45 minutes).

Limitations & Future Work

• Vulnerability to corrupted guidance inputs: When guidance contrasts exhibit head motion artifacts, extreme slice thickness, or severe anisotropy, cross-modality conditioning can transfer blur and geometric distortion into the target reconstruction.
• Substantial offline pre-training compute for ReN2N: Training conditional diffusion re-noisers and sampling multi-instance synthetic pairs requires substantial GPU resources (e.g., approximately 80 V100 GPU days for 960 OASIS scans), which presents a computational barrier for low-resource environments.
• Promising directions: Developing resolution-invariant cross-contrast attention architectures and exploring few-step diffusion distillation for the re-noising phase to accelerate offline dataset generation.

Related Work & Insights

• vs Noisier2Noise (Moran et al., CVPR 2020) & S-N2Void (Broaddus et al., ISBI 2020): Both aim for self-supervised single-image denoising, but Noisier2Noise's additive Gaussian noise alters physical noise statistics, while S-N2Void's blind spots discard vital spatial context. YADO preserves 100% of the input image and accurately mirrors real physical acquisition noise.
• vs BME-X (Sun et al., Nature Biomedical Engineering 2025): BME-X attempts general-purpose MRI enhancement using heuristic synthetic degradations and tissue segmentation losses, yet ranks among the worst on uncorrupted physical test sets (ranking behind Identity). YADO demonstrates that principled physical noise modeling yields superior stability across unseen clinical domains.
• vs BM4D (Maggioni et al., IEEE TIP 2012): BM4D remains a resilient classical benchmark but suffers from high computational latency and edge-smoothing. YADO consistently surpasses BM4D across all physical acquisition benchmarks.

Rating

• Novelty: ⭐⭐⭐⭐⭐ [Pioneers physical Noise2Noise on authentic MRI repeats and introduces diffusion-based re-noising for single-scan training]
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Extensively benchmarked across 673 participants, 14 test configurations, three major scanner vendors, and 17 competing methods]
• Writing Quality: ⭐⭐⭐⭐⭐ [Rigorous theoretical formulation directly resolving the bias of synthetic evaluation metrics]
• Value: ⭐⭐⭐⭐⭐ [Fully open-source code and model weights; matches NEx=7 acquisition quality to offer immense practical clinical value]